Support for CUDA 9.0

Add explicit __syncwarp to bias_op
 - Makes warp-synchronous code safe on Volta
Add sync mask to __shfl intrinsics
Add libdevice bytecode paths for CUDA 9
 - In CUDA 9, all supported architectures are merged into a single file
Update code gating for CUDA 9
Add sm_70 to the lookup table used by XLA
Change the default sm arch from 20 to 30.
Fix for NVPTX not yet supporting sm_70
Remove unnecessary cuda decorators from defaulted constructors
Use updated NCCL for CUDA 9 fp16 support